You are on page 1of 2

Ngo, Thuan Van November, 26th, 2015

Question 10.
(a) Why is Sokoban a good testbed for single-agent search?

Sokoban implicates its sophistication under the simple playing rules to become one of the most
challenging games to solve for single-agent search. By analyzing Sokoban in the context of applying
single-agent search, it prompts three major problems relating to lower bound, deadlock, and
branching factor [1]. For lower bound aspect, it is hard to get a tight lower bound on the solution
length for obtaining the solving efficiency. The stones can have complex interactions, with long
elaborate maneuvers often being required to reposition stones. Moreover, for some problems,
without a deep understanding of the problem and its solution, it is very difficult to get a reasonable
bound. In term of deadlock, though in most applying cases of single-agent search, all state
transitions preserve the solvability of the problem (but unnecessarily the optimality of the solution),
in Sokoban, moves can lead to states that are provably unable to lead to solutions, also known as
deadlock state. Considering the branching factor, when sliding tile puzzles and Rubik's Cube have a
maximum branching factor of four and 18 respectively, Sokoban with 34 stones, the maximum
branching factor is up to 136. We see that the branching factor surprisingly arises while the number
of stones increases. Unluckily, the large branching factor severely limits the search depth that can be
reached. Therefore, Sokoban could be the best environment to investigate the ability of single-agent
search.
(b) Why Search is one of the fundamental methods in artificial intelligence? What is the
advantage and limit?

Search plays an extremely important role in every aspect of AI. Many goal based agents are
essentially problem solving agents which must decide what to do by searching for a sequence of
actions that lead to their solutions. For production systems, we have seen the need to search for a
sequence of rule applications that lead to the required fact or action. For neural network systems, we
need to search for the set of connection weights that will result in the required input to output
mapping [2]. The advantages and limitations of search distinguish among different kinds of search
algorithms. For instances, in case of uninformed search, also called blind search, it is a class of
general purpose search algorithms that operate in a brute-force way. These algorithms can be
applied to a variety of search problems, but since they don't take into account the target problem.
For breadth-first search, its main idea is that the search tree is explored before expanding its nodes
to search their children. Therefore, it has the advantage of the simplicity in implementation but
heaviness in memory requirements. Otherwise, depth-first search drills deep into the hierarchy,
exploring one lineage until it reaches the child/leaf node with no successors. It improves on the
memory requirements by deleting entire line once explored but potentially runs into infinitely deep
search path (non-complete). As a whole, to take advantages search algorithms, we should flexibly
apply various types of search for resolving different kinds of problems.
(c) What would be real-world-like search-space property?

In general, search space in AI could means a set of all possible solutions to a problem [3]. At first, a
real-world-like search-space property is suggested in the scope of Sokoban game solving with much
concern about the large search space of this problem [4]. From that, we could understand real-

Ngo, Thuan Van November, 26th, 2015

world-like search space property is a terminology that refers to the enormous complexity of doing
searching in any case.
References

1. Sokoban: A Challenging Single-Agent Search Problem - Andreas Junghanns, Jonathan


Schaeer
2. IAI : Search John A. Bullinaria, 2005
3. Introduction to Artificial Intelligence (G51IAI): Problem Space and Search Tree - Rong Qu
4. Pushing the limits: New developments in single-agent search - Andreas Junghanns (1999)
Ph.D. Thesis, University of Alberta, Edmonton, Alberta

You might also like